Work

Patherea: Cell detection and classification for the 2020s

Year: 2025

Type: article

Abstract: • Novel point-based cell detection and classification approach that effectively utilizes multiple proposal candidates. • The largest publicly available manually labeled dataset for point-based cell detection and classification. • An improved evaluation protocol for point-based cell detection and classification. • State-of-the-art results reported on existing public and newly proposed datasets. • First to use pre-trained Transformer-based foundation models for point-based cell detection and classification. We present Patherea, a unified framework for point-based cell detection and classification in histopathology. Our method directly predicts cell locations and classes without intermediate representations and incorporates a hybrid Hungarian matching strategy for more accurate point assignment. Patherea supports flexible backbones and training regimes and leverages pathology foundation models for point-based detection, a capability not feasible with most existing point-based methods. To support method development and benchmarking, we introduce the largest fully manually labeled dataset for Ki-67 IHC, annotated by expert pathologists on full-resolution whole-slide images following standard clinical protocols. We evaluate our approach on existing public point-based datasets—Lizard, BRCA-M2C, and BCData—demonstrating superior F1 scores compared to recent point-based detection methods, while revealing performance saturation on these benchmarks. The Patherea dataset provides a complementary benchmark, capturing clinically relevant low-abundance cell classes where current approaches underperform. Finally, we identify and correct common errors in existing evaluation protocols and provide a benchmarking utility for reproducible assessment. The Patherea dataset and code are publicly released to support future research and fair comparisons.

Source: Medical Image Analysis

DOI: 10.1016/j.media.2025.103868

Authors: Dejan Štepec, Maja Jerše, Snežana Đokić, Jera Jeruc, Nina Zidar, Danijel Skočaj

Institutions: University of Ljubljana, Institute of Oncology Ljubljana

Language: English

Topic: AI in cancer detection

Subfield: Artificial Intelligence

Field: Computer Science

Domain: Physical Sciences

SDG: Good health and well-being

Open Access status: hybrid

Estimated APC paid: $4,170